Is positive_cameras a noisy version from the pivot_cameras with high IoU

[rambleramble]

Could you shed some light on how the positive_cameras related to the pivot_cameras?
It seems the pivot ones are sampled based on the parameter proposed by your paper,
and positive is a noise version of it with controlled high IoU, which is used in training the Siamese, Is this a correct understanding? If so, could you share the procedure to generate the positive_cameras from the pivot_cameras

Thank you in advance.

[lood339]

Sure, positive_cameras are cameras that are close to pivot_cameras in camera pose space. A positive camera is sampled by adding a small angular/focal length difference (generated from Gaussian distribution) to the corresponding pivot camera. As a result, the feature distance between a pivot camera and the corresponding positive camera should as small as possible.

There is a function 'sample_positive_pair' in https://github.com/lood339/SCCvSD/blob/master/python/util/synthetic_util.py to do this work.

Hope that helps.